LTX-2 is a 19B parameter AI video model that generates high-quality video clips from either text prompts (T2V) or images (I2V). It runs locally in ComfyUI and supports up to 1280x720 resolution.

What GPU do I need for LTX-2 Video?

LTX-2 was tested on an RTX 4090. For lower VRAM cards, use the fp8 model variants and enable the --novram flag. At least 12GB VRAM is recommended for reasonable generation speeds.

How to Use LTX-2 Video in ComfyUI: I2V & T2V Guide (2026)

LTX-2 lets you create realistic AI videos from text or images — all running locally inside ComfyUI. In this guide you'll find step-by-step setup instructions, six real example prompts with output videos, and solutions for common issues like out-of-memory errors.

What is LTX-2 Video?

Quick answer: LTX-2 is a 19B parameter AI video model that generates high-quality video clips from either a text prompt (T2V) or a still image + prompt (I2V). It runs locally in ComfyUI at up to 1280×720 resolution.

LTX-2 is an AI model for making high-quality videos from either text (Text to Video — T2V) or images (Image to Video — I2V). It can run locally with ComfyUI, giving you full control over video generation. Many people use LTX-2 for creating short clips, animated scenes, or video ideas that are hard to film yourself.

The model's 19B fp8 checkpoint keeps VRAM requirements manageable, and the included distilled LoRA reduces generation time significantly. Test setup used for the examples in this guide: RTX 4090, 1280×720, 20–40 second clips.

LTX-2 Image to Video (I2V) Examples

With LTX-2 I2V you start with a still image and a prompt describing the motion — the model animates it into a video clip. Below are three real examples showing the input image, output video, and the exact prompt used.

Test setup:RTX 4090 · 1280×720 · 40-second video · Prompts generated with ChatGPT

Example 1 — Image to Video

Source Image

Generated Video

LTX-2 I2V output — click play to watch

Prompt 1

A natural, conversational indoor scene. The camera holds a steady medium two-shot, framing one young woman and one young man seated across from each other at a small table, both slightly angled toward the camera with full faces clearly visible.

Both characters speak in clear, natural English with accurate lip-sync. Movements are subtle and realistic: gentle head nods, natural blinking, small hand gestures, and relaxed posture shifts while listening and responding.

Dialogue & Action Flow:

Girl (leans slightly forward, calm but enthusiastic, light hand gesture):
"Honestly, I loved the movie. The cinematography was beautiful, and the story kept me hooked the whole time."

Boy (relaxed posture, thoughtful expression, slight head tilt):
"I don't know… visually it was great, but the pacing felt slow to me, especially in the second half."

Girl (smiles, nods gently):
"That's fair, but I think the slow pace helped build the emotion. The ending really paid off."

Boy (small shrug, half-smile):
"Maybe. I did like the performances though. The lead actor really carried the film."

Girl (leans back slightly, friendly tone):
"Yeah, we can agree on that. Even if you didn't love it, it's still worth watching."

The camera remains stable throughout, maintaining consistent framing. Lighting is soft and even, the background stays softly blurred, and the scene ends naturally as both characters relax into a comfortable pause.

Example 2 — Image to Video

Source Image

Generated Video

LTX-2 I2V output — click play to watch

Prompt 2

Cinematic animated performance scene featuring a stylized puppy character, based exactly on the input image.
The puppy stands center frame on a small, colorful stage, facing the camera in a medium shot. The camera remains mostly steady with a very slow, smooth push-in throughout the video.

The puppy is singing cheerfully in clear, expressive English with accurate lip-sync that matches the lyrics. Mouth shapes are well defined for singing, with rhythmic opening and closing aligned to syllables. Facial expressions are joyful and animated, including subtle eyebrow movement, gentle blinking, and a happy, energetic smile between lyric lines.

The puppy holds a small microphone in one paw and performs light, playful gestures with the other paw while singing. Head movements are soft and rhythmic, slightly swaying to the beat.

🎵
"Hello my friends, I'm singing today,
Wagging my tail in a happy way.
Clap your paws, come sing along,
Life feels better with a cheerful song."
🎵

Example 3 — Image to Video

Source Image

Generated Video

LTX-2 I2V output — click play to watch

Prompt 3

Cinematic, natural lifestyle scene in soft daylight. The camera holds a steady medium shot of a young woman framed from the waist up, centered and facing the camera in a modern, softly lit interior. Lighting remains even and flattering, with a clean, softly blurred background.

She looks directly into the camera and begins speaking in clear, natural English with realistic lip-sync. Her facial expressions are relaxed and engaging, with subtle smiles, natural blinking, and gentle eye focus. As she talks, she casually explains her daily routine in a relatable way — waking up slowly, making time for herself in the morning, choosing comfortable clothes, and easing into the day rather than rushing.

The camera remains stable throughout, with a very slow, almost unnoticeable push-in over time to add cinematic depth. No cuts or sudden movements occur. The mood stays calm, friendly, and authentic.

LTX-2 Text to Video (T2V) Examples

LTX-2 T2V generates video entirely from a text prompt — no input image required. It's fast, flexible, and great for bringing scenes to life from scratch.

Test setup:20-second video · default workflow settings · Prompts generated with ChatGPT

Example 1 — Text to Video

Generated Video

LTX-2 T2V output — click play to watch

Prompt 1

A colorful, imaginative performance scene featuring a cheerful vintage-style puppet musician made of soft felt and stitched fabric. The puppet has round expressive eyes, a wide stitched smile, and short woolen hair, wearing a quirky outfit with suspenders and a bow tie. It stands center frame on a small playful stage, holding a classic handheld microphone.

The camera opens in a medium, front-facing shot. Bright, warm lighting fills the scene, creating a fun, upbeat atmosphere with soft highlights on the puppet's fabric texture.

The puppet begins singing joyfully in clear English, its fabric mouth opening and closing rhythmically with accurate lip-sync.

🎵 "Sing it loud, feel the day, Happy tunes will find their way, Clap your hands, move along, Life feels better with a song!" 🎵

Example 2 — Text to Video

Generated Video

LTX-2 T2V output — click play to watch

Prompt 2

A vibrant, futuristic workshop filled with bright neon accents, glowing holographic panels, and colorful cables hanging from the ceiling. Warm pink, teal, and yellow lights illuminate the space, creating an energetic and creative atmosphere.

A young woman engineer wearing a sleek, translucent visor stands up from a workbench after finishing a small robotic gadget. From an upper platform, a friendly alien inventor with colorful skin and glowing accessories slides down a ramp, holding a floating, glowing cube that shifts colors as it spins.

She laughs softly, eyes widening with excitement, and replies: "Tell me you didn't just invent something amazing again."

Example 3 — Text to Video

Generated Video

LTX-2 T2V output — click play to watch

Prompt 3

A bright, high-energy football stadium during a live match. The crowd fills the background, cheering loudly, flags waving, sunlight reflecting off the green pitch. The camera opens in a steady medium two-shot of a young football player boy and a confident, athletic girl standing near the sideline.

The boy says: "We keep pushing down the left wing. Their defense is breaking every time."

She replies calmly: "Yeah, but the real opening is after the second pass. Draw them in… then cut straight through the middle."

Natural dialogue, realistic lip-sync, subtle head movement and gestures, bright daylight, shallow depth of field, cinematic sports atmosphere.

How to Set Up LTX-2 Video in ComfyUI

Follow these three steps to get LTX-2 running inside ComfyUI.

Step 1 — Update ComfyUI

Go to ComfyUI_windows_portable\update and run update_comfyui to get the latest version.
Start ComfyUI, click Manager, then Switch ComfyUI, and select the newest version.
Click Update ComfyUI to update all installed custom nodes as well.

Step 2 — Download Required Models

You need four model files — download each and place them in the correct folder:

Checkpoint: ltx-2-19b-dev-fp8.safetensors
Text Encoder: gemma_3_12B_it_fp8_e4m3fn.safetensors
LoRA (distilled): ltx-2-19b-distilled-lora-384.safetensors
LoRA (IC detailer): ltx-2-19b-ic-lora-detailer.safetensors
Spatial Upscaler: ltx-2-spatial-upscaler-x2-1.0.safetensors

Place each file in the correct folder:

ComfyUI/
└── models/
    ├── checkpoints/
    │   └── ltx-2-19b-dev-fp8.safetensors
    ├── text_encoders/
    │   └── gemma_3_12B_it_fp8_e4m3fn.safetensors
    ├── loras/
    │   ├── ltx-2-19b-distilled-lora-384.safetensors
    │   └── ltx-2-19b-ic-lora-detailer.safetensors
    └── latent_upscale_models/
        └── ltx-2-spatial-upscaler-x2-1.0.safetensors

Step 3 — Load the Workflow in ComfyUI

Open ComfyUI and drag-and-drop the downloaded workflow JSON file onto the canvas to load it.
In the subnode graph, confirm CKPT_name is set to ltx-2-19b-dev-fp8.safetensors.
Confirm the LTXV audio text encoder loader node is set to gemma_3_12B_it_fp8_e4m3fn.safetensors.
For I2V: add your source image. For T2V: just add your prompt. Then click Queue Prompt to run.

Fixing Out of Memory (OOM) Errors

Running out of GPU memory is the most common issue with LTX-2. Here are three reliable fixes, in order of priority:

Update NVIDIA Drivers

Make sure your graphics drivers are up to date. Visit the NVIDIA driver download page and install the latest version for your GPU.

Add the --novram Flag

Find run_nvidia_gpu.bat (or similar) in your ComfyUI folder. Right-click → Edit with Notepad. Change the command to:

.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --novram
pause

Save the file and restart your ComfyUI server.

Increase Windows Page File Size

If memory problems continue after the steps above, increase your system's virtual memory (page file) size. This video explains how: How to change Page File Size in Windows. After changing the setting, restart your computer. Many users find this resolves persistent OOM errors.

💡 Tip: Always start with the fp8 checkpoint rather than any full-precision variant. The fp8 model uses roughly half the VRAM and is what all examples in this guide were generated with.

Tips & Troubleshooting

Video looks strange or incoherent

Fix: Use more specific, detailed prompts. Describe camera movement, lighting, character actions, and scene mood explicitly. Shorter, clearer prompts often give better results than very long ones.

Out of memory (OOM) errors

Fix: Follow the three-step OOM fix above: update drivers, add --novram flag, increase page file. Try lowering resolution or step count first before other changes.

Model files not appearing in dropdowns

Fix: Make sure all files are in the correct folders. Press F5 or Ctrl+Shift+R to fully refresh ComfyUI in your browser after adding new files.

ComfyUI "nightly" build bugs

Fix: Nightly builds get new features faster but may have small bugs. Update ComfyUI frequently. If something breaks, try switching back to the previous stable version via the Manager.

⚠️ Warning: Always make sure all five model files are placed in the correct folders before running the workflow. Missing a single file is the most common cause of workflow errors.

Hope this helps you get started with LTX-2 AI Video and ComfyUI! Feel free to share your results or ask questions in the comments below.

Frequently Asked Questions

LTX-2 is a 19B parameter AI video model that generates high-quality video clips from text prompts (T2V) or images (I2V). It runs locally in ComfyUI and supports up to 1280×720 resolution.

You need: ltx-2-19b-dev-fp8.safetensors (checkpoint), gemma_3_12B_it_fp8_e4m3fn.safetensors (text encoder), two LoRA files (distilled and IC detailer), and the spatial upscaler. All are free on Hugging Face.

Update your NVIDIA drivers, add the --novram flag to your ComfyUI launch script, and if needed increase your Windows page file size. These three steps resolve most OOM errors.

I2V (Image to Video) animates a still image using a text prompt to guide the motion. T2V (Text to Video) generates video entirely from a text description, without requiring an input image.

LTX-2 examples in this guide were generated on an RTX 4090. For lower VRAM cards, use the fp8 model and enable --novram. At least 12GB VRAM is recommended for reasonable generation speeds.